基於WiFi 通道時間空間特徵之多人姿態辨識技術

Abstract

由於WiFi裝置的普及性，基於WiFi CSI的姿態估測技術獲得廣大的討論，然而現有的研究大部分都只應用於單人的情境，且現存的多人應用系統也因使用CSI的振幅與相位當作輸入而使系統在環境與時間變動下表現不夠穩定。於此，本論文探討使用CSI中的空間特徵，如信號入射角(AoA)與飛時測距(ToF)，與時間資訊中的都卜勒頻率位移(Doppler shift)，來達到多人姿態估測，這些特徵能夠捕捉使用者的位置與動作資訊，並不會受到環境變動影響。現存的特徵擷取方式只在時間軸與頻率軸擇一來做CSI信號處理，大大刪減OFDM可以達到的偵測與抗噪能力。於此，我們提出使用MUSIC演算法來沿著時間軸與頻率軸處理CSI，達到AoA-ToF與AoA-Doppler的估測，我們也提出用MVDR波束成行器(beamformer)來增加AoA-Doppler估測的多樣性。我們搭建了一個模擬平台來方便驗證我們的特徵擷取方法是否有效，在模擬中，我們的方法相比於文獻，更能壓抑雜訊、干擾信號和增加解析度。我們也設立真實的實驗平台並設計兩種實驗：同場域實驗與跨場域實驗。在同場域中，我們特徵擷取方式相較於文獻中的方法，姿態辨識準確率高了20%，而在跨場域中，我們的方法與文獻中使用CSI振幅與相位的方法相比，姿態辨識準確率提高了12%，證明了我們提出的時間空間特徵能應用於多樣的環境中。

Pose estimation based on WiFi CSI receives skyrocketing popularity as WiFi devices are ubiquitous. However, most existing works only focus on single-user applications, and the existing multi-user pose estimation systems are not adequately robust since they utilize CSI amplitude and phase which vary in different CSI collection timings and environments. Thus, in this thesis, we focus on spatial features, such as angle of arrival (AoA) and time of flight (ToF), and temporal features, such as Doppler shift, embedded in WiFi CSI for multi-user pose estimation. These features capture position and motion information that is uncorrelated with data collection timings and environments. Existing feature extraction methods take either only CSI data along the time axis or only CSI of different OFDM subcarriers into consideration, limiting the feature resolutions and denoising ability. To handle this problem, we propose a MUSIC-based method to jointly estimate AoA-ToF and AoA-Doppler pairs from consecutive CSI samples and different subcarriers. We further propose to use an MVDR beamformer to extract AoA-Doppler pairs to increase the resolution of Doppler shifts. A simulation platform is deployed to generate CSI with controllable AoAs, ToFs, and Doppler shifts for convenient verification. In the simulation, our proposed methods can obviously suppress noises and increase the resolutions of desired features compared to the existing methods. We also set up a platform to collect real data. Two experiments are done in the same-domain scenario and the cross-domain scenario. In the same-domain scenario, our feature extraction methods outperform the existing methods by 20% precision, showing that our methods can suppress noises and increase the feature resolutions. In the cross-domain scenario, our system outperforms the existing model using amplitude and phase as input by 12% precision, demonstrating that our spatial and temporal features are more generalized.